alea-institute/alea-multihop-qa

ALEA flan2: multihop_qa

Each row gives a real public-record passage and a question that needs two or three linked facts from different sentences of the passage (find an intermediate fact, then use it). Responses give the direct answer, short worked steps quoting the passage and ending in "Answer: ...", or the intermediate and the final answer.

Kinds: mh_direct, mh_steps, mh_bridge (intermediate fact + answer); sub gives 2hop / 3hop.

Sources

Passages (250-1,100 words) come from public-record KL3M data: 10-K annual reports (MD&A sections), U.S. government web pages (guidance, news, FAQs), congressional committee reports, material contracts filed as EDGAR exhibits, congressional hearing transcripts, state and federal court opinions (Caselaw Access Project), federal district court orders and opinions (RECAP), Federal Register notices, Federal Register proposed and final rules. No passage was edited, except for the deliberate twin variants described below, which are labelled as such in the gold record. Names of parties, judges, officials and companies are public record and are kept. Passages with private e-mail addresses or phone numbers were dropped. No more than 30% of documents come from one source pool, and no more than 5% from one court, agency, filer or host.

How the gold was made and verified

• Questions and their hop chains (sub-question, answer span, quote for each hop) were written by GLM-5.3-Flash from the passage.
• Code checks: every quote and span occurs verbatim in the passage, each hop answer lies inside its quote, the final answer is the last hop's answer, hops use different sentences (nupunkt), the intermediate (bridge) answers do not appear in the question, and the question shares few content words with the final-answer sentence.
• A second model from a different family (DeepSeek-V4.1-Flash) answered each question blind from the passage; items were kept only on agreement.
• One-hop shortcut test: the same model answered again from the passage with the bridge sentence(s) removed. Items it could still answer were dropped as one-hop questions in disguise (most written questions failed this test; the release is small on purpose).
• Rows that show the intermediate fact (steps / bridge) also required the solver to answer the first sub-question blind with the same intermediate answer; three-hop items are rendered as direct answers only.
• Compound questions and questions about e-mail addresses were dropped.
• Worked steps quote the passage verbatim and end with "Answer: ...".

Rows that failed any check were dropped, never patched.

Size and composition

1,477 training rows from 1,319 source passages (1,610,203 tokens in the ALEA sw32k tokenizer). Presentation: plain instruction, Q/A frame, or chat turn (instruction or user turn always first).

kind	rows	share
mh_direct	851	57.6%
mh_steps	470	31.8%
mh_bridge	156	10.6%

source pool	rows	share
10k	480	32.5%
dotgov	267	18.1%
crpt	226	15.3%
agr	173	11.7%
chrg	158	10.7%
cap	103	7.0%
recap2	36	2.4%
fr	31	2.1%
frrule	3	0.2%

Checks against shortcuts

Bag-of-n-gram baselines that see only part of the input (cross-validated by source passage), compared with the majority-class prior:

{
 "question_only": null,
 "note": "All encoded rows are answerable (the removed-sentence abstention rows were disabled after blind review r1), so there is no class label for a question-only classifier. Shortcut control is per item: a blind DS solve on the passage with the bridge sentence(s) removed must fail (5,740 of 7,333 solver-agreed items dropped as one-hop in disguise), plus a code check that the final-answer sentence does not hold >= 75% of the question's content words."
}

Decontamination and review

Rows sharing long n-grams with evaluation items or held-out sets were removed; overlaps made only of document boilerplate (headers and forms common to many source documents) were logged and kept. Held-out documents (about 1.5% of source documents) are not in this release. Builder spot checks read 39 rows against their sources. Independent blind reviews of 60 rows each found problems in round 1 (mostly in abstention rows built by removing a sentence, which were then removed from the family, plus one wrong intermediate fact, which led to the extra intermediate check) and 3 wrong rows of 60 in round 2 (two compound questions answered in part and one e-mail placeholder); questions of those two kinds were then removed.

Cost

LLM spend for this family: about $8.48.